Wire manufacturing: when AI creates the quotation

Source: worldsteel Association

Wire manufacturers know: Quotation preparation is usually a time- and resource-intensive process, sometimes taking up to two weeks. This is because a calculation including all process parameters is often based to a large extent on individual expert knowledge, which always leaves room for inaccurate results. A new, more efficient way is to generate quotations using machine learning (ML). The ML model automates the calculation of a process chain in wire production and determines the optimal sequence of steps in the production process more reliably than humans.

Drawing plant for wire production (Source: C. Wagner)

Wire production is a complex process. The path from raw material to finished wire involves many small steps. First, the raw wire is drawn through special tools and brought to its target diameter. Precise and controlled work is required to guarantee the desired tolerance and surface quality of the wire. The wire is then treated with heat - "annealing" is intended to improve its mechanical properties. Depending on the subsequent product application, the wire is also given a protective coating, for example to protect it from corrosion or to insulate it electrically. Once the wire has finally been drawn and processed according to the desired customer-specific requirements, it is wound onto spools or reels and shipped.

In view of the enormous process complexity in wire production, precise quotation preparation optimized in terms of process management is of particular importance. In addition, increasing price pressure and limited personnel resources in global competition are increasingly challenging the often medium-sized contract manufacturers of customized products to use their limited resources as efficiently as possible. In particular, the preparation of quotations and the associated quotation costing are regarded as resource-intensive business processes that often do not even lead to actual orders in the end.

In the case of elaborate quotations, these can result in costs per quotation of over €1,000. Assuming 1,000 offers per year and realistic offer success rates of often less than 30%, this means costs averaging around €700,000 per year.

The input parameters that the customer provides to the sales department are mostly manageable: grade, thickness, width, length, melting, surface quality, finishing, final condition. Nevertheless, the preparation of a quotation requires a great deal of effort, combined with a high level of expertise. The right process parameters and finally an entire route have to be worked out. The wire expert must prepare this data, i.e. find and correct inaccurate data, convert it into numerical data and apply unique designations to normalize the data.

With all this, the problem cannot be represented by a classical classification model because it depends on the process flow.

So how can machine learning (ML) and intelligent algorithms specifically help optimize a wire manufacturer's highly customized quoting process here?

Analysis of numerous data by means of ML

Since the quotation price for the customer is mostly derived from the preliminary planning of the production process, the work plan, the ML software was first fed with the wire manufacturer's work plans: the training basis for the system included around 850 work plans with a minimum of five, an average of eleven and a maximum of 28 production steps. In addition, typical process parameters such as the planned production time, the number of annealing bricks and the annealing temperature were included. The complexity of the choices is shown schematically in the process mining diagram (See Figure 1). This production system refers to actual data from hundreds of work plans.

Fig. 1: Process Mining Diagram of wire production (Source: C. Wagner)

Using process mining approaches, process routes could now be clustered and the solution space for model building reduced. Based on a decision tree procedure, a model was then developed which, with the help of the customer specification, derived an ideal work plan based on which, a target price was automatically calculated.

Based on the test data, a model quality of over 95% was finally achieved, i.e. the automatically estimated costing-relevant parameters matched the human planning with a high degree of accuracy.

ML creates the optimal process route

So how does the automated model go about creating the routing? First of all, it decides whether it is a rough or a fine train. This already significantly limits the selection of possible routes. The decision is based on the required draw ratio, i.e. the ratio of the initial to the final diameter. Secondly, a decision is made on the feedstock. This must have a next larger diameter and be of the same quality. In the third step, the model predicts a binary code for each machine and thus determines whether it will be used. In the fourth and final step, the model determines the process parameters relevant for the calculation.

Supervised machine learning methods were used to create a handful of models: Generalized Linear Model, Deep Learning, Decision Tree, Random Forest, Gradient Boosted Trees and Support Vector Machine. For the specific example of 1.6 mm wire to 0.45 mm wire, the decision tree demonstrated the best relationship of model quality to computation time and understandability. Figure 3 shows this tree. There you can see that the model predicted the correct path with an accuracy of about 98% with only two parameters.

Fig. 2: Step-by-step concept for forecasting the work plan (Source: C. Wagner)

Fig. 3: Decision tree and model accuracy of the forecast (Source: C. Wagner)

Once decision trees are established for all parameters, the next step is to integrate them into the enterprise software. The cost rates are statistically stored in the ERP system and sales order information remains stored in a quotation mask as before. A REST API interface is used to transfer the data to a server when it is saved. The trained models then return the corresponding forecast parameters. When offset against the cost rates, the result is a value that the company can convert into a target price quotation.

Of course, this does not produce a perfect result. The model quality is approximately 90 percent. To what extent this is sufficient for a price quotation is ultimately left to the sales managers in the company. It is not as important to miss the annealing temperature by a few degrees, for example, as to plan a completely wrong route instead. Therefore, it is recommended to check a quotation forecast in advance by the work preparation department. However, compared to a complete redesign of a quotation, this effort is significantly lower - on average about 80 percent compared to the conventional approach.

For Prof. Dr.-Ing. Carsten Wagner, who initiated the ML project with his team as part of a research project at the Göttingen University of Applied Sciences and Arts (www.hawk.de) and accompanied the complete automation of the quotation process via ML, the time factor, in addition to saving resources, is also the decisive factor for the wire manufacturer to now be best equipped "quotation-wise" for global competition: "Reducing the quotation preparation time from 2 weeks to the push of a button in extreme cases - that's really disruptive!"